Let a genome decide, and report where each axis actually acts

M3.6 joins the M3 pieces: strategy_for derives a StrategyConfig from the
metabolic budget, founders are pre-diverged variants admitted to the lineage
archive, and evo-sim --ensemble runs them over a resource and temperature grid
with every cell's seeds split in half and never merged. ADR-0029's driver is
untouched, so the M2.6 fingerprint still holds. Nothing mutates during an
episode, so NPC mutation stays open.

Three things measurement found, kept rather than smoothed over. The report
aggregates were default-initialized, leaving their scalar counters
indeterminate. The resource level never reaches strategy derivation at all, so
the header now says where each axis acts instead of claiming both shape a
strategy. And the resource axis is not monotonic: a sweep peaks near 8000
femtograms and falls away on both sides, because only one consumer may bite a
target per tick, so scarcity concentrates mass into individuals that reach the
threshold while abundance spreads the same income too thin. The test asserts
that an axis moves the outcome and never a direction; an earlier draft asserted
a direction that held at two cells and the sweep disproves.

Earning a birth at all required pinning the investment fraction to the schema
floor, since intake scales with surface and investment with mass. That is a
finding about the ADR-0031/0032 cost model, now the second priority in STATUS.

Release 33/33, analysis, ASan 33/33 and clang-format all ran and passed here;
no Clang or GCC compile happens on this host, so Linux CI is still the first
result for this increment.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
dingo
2026-09-23 23:03:52 -04:00
co-authored by Claude Opus 5
parent fdaed3e4f5
commit 89ad95bf12
14 changed files with 1063 additions and 12 deletions
+76
View File
@@ -720,3 +720,79 @@ migration are M7 and M8, and NPC mutation remains undecided. Golden
`d61ad5bf0c928c76`, ecosystem fingerprint `2cd4a15f9fbe9949` and the frozen genome
fixture are unchanged. No deployment manifest or image pin changed. Linux
validation is pending; no Gitea result is claimed since `b01a824`.
## M3.6 genome ecology ensembles
Executed on this Windows host (2026-09-23): Release configure, build and `ctest`
33/33; the analysis preset (clang-tidy 19, warnings as errors) built clean; the
ASan preset configured, built and ran 33/33; `clang-format` 19.1.5 reported no
violations after the new files were formatted. All four checks ran and passed on
the committed tree. No Clang or GCC compile happens on this host, so the standing
gap recorded for `29ce255` is unchanged and Linux CI remains the first Clang and
GCC result for this increment.
`evo-ensemble-tests` covers eight properties: that a derived strategy carries
exactly ADR-0032's budget quantities and invents no margin, that divergence is
reproducible and refuses its bounds rather than truncating, that conservation
holds at every tick of every cell, that a run is reproducible and a different seed
is not, that the archive ends holding exactly the founders and every organism that
lived with no living carriers, that the two seed halves agree within the stated
band, that each gradient axis alone changes the outcome, and that a thermally
tolerant strategy keeps more scheduled life at the extremes. `evo-sim --ensemble`
is pinned by fingerprint `c2b96c86101be628`, confirmed stable across three runs
and distinct from a different seed's `8f88e72cf2715eac`.
Three defects were found and fixed during implementation, all by measurement
rather than by review. `EnsembleReport` and `CellReport` were default-initialized
rather than value-initialized: both are aggregates whose vector members give them
a non-trivial default constructor, which leaves the scalar counters indeterminate,
and the episode total read as garbage. Both are now `{}`.
The other two are findings about the model, and are recorded in ADR-0035 rather
than tuned away. The header claimed the resource level entered both the derived
strategies and the episode's nutrient pool. It does not enter the derivation at
all: founders are derived fully grown, so their construction transition is empty
and ADR-0031's per-channel `available` cannot reach intake, assimilation or
reproduction investment. The comment was corrected rather than the code, because
the code was right and the claim was false.
The resource axis was then measured to be decorative at the original constants and
non-monotonic once it was not. A ten-point sweep at the original fixture returned
41 to 45 organisms across three orders of magnitude of nutrients with identical
tick totals and eight extinctions everywhere: every birth came from the opening
endowment, because a body of 353,600,000 fg needed 73,776,000 fg of carbon to
reproduce while feeding supplied 8,000 fg per tick before assimilation, or roughly
23,000 ticks against a 60-tick lifespan. Rescaling so a birth can be earned
required pinning the investment fraction to the genome schema's floor of one per
thousand, because intake scales with surface and investment with body mass, so
the ticks needed to fund a birth fall only as the cube root of mass. That a
reproduction is affordable only at the schema floor is a finding about the
ADR-0031/0032 cost model and is now the second priority in STATUS.md.
After rescaling, a sweep of 200 to 200,000 femtograms gave 45, 45, 45, 51, 61, 81,
44, 44, 42 and 42 organisms: a peak near 8,000 and a fall away on both sides. The
mechanism is ADR-0026's feeding rule that a target may be bitten by only one
consumer per tick, which caps community income at one producer bite however large
the pool; scarcity empties the producer, consumers bite each other, and that
concentrates mass into individuals that reach the threshold, while abundance
spreads the same income too thin to fund a birth. The gradient test therefore
asserts only that an axis moves the outcome, never a direction. An earlier draft
of that test asserted `rich_organisms > lean_organisms`; that assertion held at
the two chosen cells but implied a monotonicity the sweep disproves, so it was
replaced rather than satisfied by picking two cells that flattered it.
Coverage limits. The grid is four cells and the tests use eight seeds per cell, so
every number describes this fixture and nothing wider. The split-half band
(`high * 2 <= low * 3 + 2`) is a chosen tolerance, not a derived one. No organism
mutates during an episode, so nothing here exercises in-episode divergence and NPC
mutation stays open. The thermal penalty lives in the harness and no contract
knows about it, so no contract test covers it. Nothing is serialized.
The eleven new CLI tests are new-behaviour tests for a new mode, not regressions
of a fixed defect; only the M2.6 bidirectional mode check they extend was ever a
regression. They were confirmed to fail the run, which is what `WILL_FAIL`
asserts, but each one was not separately confirmed to fail through the branch it
names, so a rejection arriving from a different branch would still pass. Golden `d61ad5bf0c928c76`, ecosystem fingerprint `2cd4a15f9fbe9949` and the
frozen genome fixture are unchanged, and ADR-0029's episode driver was not
touched. No deployment manifest or image pin changed. Linux validation is pending;
no Gitea result is claimed since `b01a824`.